HansKristian-Work/vkd3d-proton v3.1

Version 3.1

on GitHub

This release is mostly just tagged now to facilitate bumping the build version shortly after to 3.2.
A ton of important changes will happen for 3.2, so this marks the last "stable" release before plunging into
uncharted territory.

Features

• Add VK_EXT_descriptor_heap support. This has been supported as an opt-in feature for many months now and has been battle tested quite a lot on NVIDIA.
  It is unfortunately not enabled by default in this release due to driver quality concerns.
  (This will change imminently on master however, which is one of the reasons this release is happening now.)
• Add support for KHR_opacity_micromap and remove old and broken EXT path. Expose DXR 1.2 when supported.
• Add support for Shader Model 6.9 which includes Shader Execution Reordering and long vector.
  Currently only lights up on NVIDIA since SM 6.9 is a very RT/ML-centric update.
• Add support for various NVAPI intrinsics in dxil-spirv.
• Expose AgilitySDK 619 features.
• Expose view instancing by default as more games rely on this.
  It's not a fully conformant implementation (but let's be real, no implementation even on Windows is ...),
  so we have an opt-out for certain games that try to push the feature to its outer limits.

Performance

• Implement an optimized path for ClearUAV. It now goes through vkCmdClearColorImage or vkCmdFillBuffer when possible.
  This allows implementations to optimize the clear in some cases.
• Optimize some pathological COPY -> RESOURCE -> COPY barrier patterns. Cannot trust applications to write competent code :')
• Avoid some compositor roundtrips in IDXGISwapChain::Present() related to occlusion checks. Should improve performance
  when CPU bound on the main rendering thread.
• Batch more RTAS build commands before inserting barriers. Massively increases RT performance in some cases.

Fixes

• Steam Frame is now a supported platform, so various fixes catering to QCOM hardware have been added, and aarch64 builds are supported.
• Handle D24 <-> buffer copies correctly on AMD GPUs where we emulate it as D32. This fixed some jank in e.g. MGS4 remaster.
• Various fixes for low_latency2 paths submitted by NVIDIA to match spec version 3 update.
• Fix some UB behavior around udiv/umod/sdiv/srem. NVIDIA compilers take advantage of the allowed UB in some cases which broke things.
• Fix undocumented render pass API resolve behavior.
• Update tight alignment behavior to match AgilitySDK 619.5 behavior.
• Massively simplify the AMD_anti_lag implementation and fixes a lot of bugs with frame attribution.
• ... A ton of random dxil-spirv fixes too numerous to list. These problems will soon go away ...

Workarounds

• Add various workarounds to make it easier to use AMD GCN hardware and QCOM hardware with games that assumes Wave32 is supported.
  A lot of games have started to simply break in half recently due to games not caring about these GPUs.
  Most of the time this can be patched in.
• Add more detailed UE5 / RE Engine detection which is intended to aid Mesa drivers perform some more annoying workarounds
  which cannot be tied directly to vkd3d-proton detection.
• Workaround broken root SRV/UAV behavior which impacts NV. Games can get away with out-of-spec behavior on Windows which we didn't replicate.
  Should remove some glitches and crashes on those GPUs in the wild.
• Rewrite how the use-after-free workarounds are implemented. Works around lots of broken behavior around FSR4 UE plugin.
• ... A ton of random game workarounds too numerous to list.